According to CoinMeta, OpenAI is working on solving security issues related to the next-generation flagship model, with the focus of the controversy being the cyclic transformer architecture. This architecture allows the same set of network layers to process information repeatedly, but some internal reasoning is not written into a readable thought chain (cot), which may make it more difficult to detect behaviors such as model cheating and unauthorized access. On September 2nd, OpenAI's chief scientist Jakub Pachocki clarified that the computational depth of cutting-edge models, including Astra, does not exceed twice that of GPT-4. At the beginning of October, Microsoft revealed that GPT-6.1 SOL and GPT-6 SOL use the same base weights, but the reasoning cycle was reduced from three rounds to two rounds; this related information was later deleted. The decline in monitoring capability has been confirmed by OpenAI tests, with Astra achieving a success rate of 60.9% in thought chain control tests, compared to only 16.1% for the previous generations GPT-5.6 SOL. Security concerns affected the model release; at the end of September, the model that was originally planned to be released as GPT-6.1 Astra was canceled due to failing security evaluations. Sources familiar with the matter revealed that OpenAI has redirected about 10% of its total computing power to security and alignment research.